Surveys show patients are cautiously open to AI in healthcare but do not trust it on its own terms. Most say they want a doctor to stay in charge, and many doubt health systems will use AI responsibly. Trust in medical AI tends to follow trust in the people and institutions using it, which means clinicians and hospitals, not algorithms, carry most of the burden of earning it.
Ask the public whether they trust AI in healthcare and you get a confusing answer. Many are excited about faster diagnosis and fewer errors. Many are uneasy about machines making decisions about their bodies. The same person can hold both views before lunch.
Read the major surveys side by side, though, and a clearer pattern emerges. Patients do not really have an opinion about “AI” in the abstract. They have an opinion about who is using it, on them, and whether that person or institution is looking out for them. Trust in medical AI is borrowed. It comes from the clinician and the health system, and it can be spent.
That is the argument of this piece, with one important complication the evidence forces us to add: simply telling patients “a doctor supervised this” is not enough to transfer trust. The borrowing works through relationships and transparency, not labels.
How much do patients trust AI in healthcare?
The best-known US data point comes from the Pew Research Center, which surveyed more than 11,000 adults in December 2022. Sixty percent said they would be uncomfortable if their own healthcare provider relied on AI to diagnose disease and recommend treatment. Only 38% thought AI would lead to better health outcomes; 33% thought it would lead to worse outcomes and 27% expected little difference. Women were more uncomfortable than men. Pew also found that a majority, 57%, expected greater use of AI for diagnosis and treatment to make the relationship between patients and providers worse. That is the fear underneath the discomfort. People are less worried that the machine will be stupid than that it will stand between them and the person who is supposed to be looking after them. Keep that in mind as you read the rest of the evidence, because it explains most of it.
A newer, international picture is somewhat warmer. The COMFORT study, published in JAMA Network Open in 2025 by Busch and colleagues, surveyed 13,806 hospital patients across 74 hospitals in 43 countries. A majority, 57.6%, had a positive view of AI in healthcare. But fewer than half expressed trust on specific items about AI’s reliability, with the lowest trust for AI information about how they would respond to treatment. Patients also showed strong preferences about how AI should be used: 70.2% preferred AI that could explain its decisions, and 72.9% preferred decisions led by physicians, even if that meant slightly lower accuracy.
Read those two numbers again. Most patients would accept a less accurate system if it meant a doctor stayed in charge. That is not a technology preference. It is a statement about accountability.
Do patients trust their health system to use AI responsibly?
Many do not. A study by Paige Nong and Jodyn Platt, published in JAMA Network Open in February 2025, used a nationally representative US survey conducted in mid-2023. Most respondents, 65.8%, reported low trust that their healthcare system would use AI responsibly, and 57.7% reported low trust that their system would make sure an AI tool would not harm them.
The most revealing finding was about who trusted. People with higher general trust in their healthcare system were more likely to believe it would use AI responsibly and protect them from harm. In other words, trust in AI was largely an extension of trust in the institution. People who felt their hospital was on their side extended that to its technology. People who did not, did not.
This is what “borrowed trust” means in practice. Health systems cannot build trust in AI separately from trust in themselves.
Does a doctor’s involvement make patients trust AI advice more?
Here the evidence adds a twist.
In a pair of preregistered experiments published in Nature Medicine in 2024, Moritz Reis, Florian Reis and Wilfried Kunde showed 2,280 participants the same medical advice with different labels: written by a human doctor, by AI, or by a doctor with AI assistance. Advice labelled as involving AI was rated as less reliable and less empathetic, and people were less willing to follow it. Crucially, this held even when participants were told a doctor supervised the AI.
So the borrowing is not automatic. Adding “reviewed by a physician” to a chatbot answer does not transfer the doctor’s credibility. If anything, the label signals that a machine was involved, and people discount the advice.
How does that fit with the COMFORT finding that patients want physician-led decisions? The two studies are measuring different things. COMFORT asked what patients want the system to look like. Reis and colleagues measured how people react to a single piece of advice from someone they do not know. The trust patients extend comes from the relationship with their own clinician, over time, not from a line of small print. A stranger’s endorsement of an AI output is worth little. Your own doctor explaining how they use a tool and why may be worth a great deal.
A worked example: two ways to introduce an AI tool
This example is hypothetical.
Two clinics adopt the same AI tool that drafts replies to patient portal messages. We have written before about the risks when doctors simply agree with AI drafts, but here the question is trust.
Clinic A adds a line to every reply: “This message was generated with the help of AI and reviewed by your care team.” It says nothing else. Patients start wondering whether their doctor read their message at all. Some complain. A few stop using the portal and phone instead.
Clinic B takes a different route. Clinicians tell patients at visits, in their own words, that the clinic uses a tool to draft some messages, that a named clinician reads and edits every reply, and that patients can always ask for a call. A short notice in the portal explains what the tool does and does not do. When a reply matters, such as a result that needs action, the clinician writes it themselves.
The technology is identical. Clinic B has spent some of its existing relationship capital to introduce the tool, with transparency about how it works. Clinic A has attached a label and hoped for the best. The Reis experiments suggest Clinic A’s approach may actively reduce trust. The Nong and Platt findings suggest Clinic B’s approach works best where trust already exists.
The strongest counterargument: trust should follow performance, not people
A reasonable objection is that all this talk about relationships is sentimental. If an AI tool is more accurate than a doctor at a narrow task, patients should trust it more, and designing systems around feelings could hold back safer care.
There is real force here. Patients who prefer physician-led decisions at slightly lower accuracy, as COMFORT found, are making a trade-off that could, in some cases, cost them. And trust built on relationships can be misplaced: a warm clinician can still be wrong.
But the objection underestimates two things. First, patients rarely have any way to verify an AI tool’s performance themselves. In the absence of evidence they can check, trusting the people accountable for the tool is rational, not sentimental. Second, trust affects outcomes. Patients who do not trust advice are less likely to follow it, as the Reis experiments showed directly. A more accurate tool that patients ignore can produce worse results than a less accurate one they act on.
The resolution is not to choose between performance and relationships. It is to publish performance clearly, so that trust has something solid to rest on, and to introduce tools through the relationships patients already value. Both matter, and bias in performance, as we discussed in our piece on algorithmic bias, can quickly destroy trust in groups that already have reasons to be wary.
What this means for health systems, clinicians and patients
For health systems. Trust in your AI will track trust in you. Invest in explaining which tools you use, what they do, how they are checked and how patients can opt for a human. Surveys suggest women and people in poorer health are more cautious; design communication with them in mind rather than as an afterthought. When tools fail, say so.
For clinicians. Your endorsement is the most valuable trust asset in the system, and it is finite. Explain in your own words how AI fits into your work, and only vouch for tools you understand. Ambient scribes and other background tools are a good test case: as we argued about AI scribes and burnout, their value depends on how the time saved is used, including more attention to the patient in front of you.
For patients. It is reasonable to ask your clinician whether AI is used in your care, what it does and who checks it. You can ask for a human review of anything important. And remember that AI tools, however sophisticated, do not replace the judgment of a clinician who knows you.
Frequently asked questions
Do patients trust AI in healthcare?
Partly. A 2025 international survey of 13,806 hospital patients found 57.6% had a positive view of AI in healthcare, but fewer than half trusted specific aspects of its reliability. In the US, Pew found in 2022 that 60% would be uncomfortable with their provider relying on AI. Most patients want doctors to remain in charge of decisions.
Why are patients wary of AI in medicine?
Common concerns include accuracy, privacy, loss of the personal relationship with clinicians, and uncertainty about who is responsible when things go wrong. A 2025 JAMA Network Open study found about two-thirds of US adults had low trust that their health system would use AI responsibly, and those with lower trust in their health system overall were more sceptical.
Does it help if a doctor reviews AI advice?
Not automatically. A 2024 Nature Medicine study found advice labelled as involving AI was rated less reliable and empathetic, even when participants were told a doctor supervised it. What seems to build trust is a patient’s existing relationship with their own clinician, plus clear explanation of how AI is used, rather than a label.
Would patients accept less accurate AI if doctors stay in charge?
Many would. In the COMFORT survey across 43 countries, 72.9% of patients preferred physician-led decision-making and 70.2% preferred explainable AI, even at the cost of slightly lower accuracy. This suggests patients value accountability and understanding as much as raw performance.
How can hospitals build patient trust in AI?
Be transparent about which tools are used and why, publish how they perform and are monitored, keep clinicians clearly accountable, offer a human alternative, and introduce tools through clinicians patients already know. Survey evidence suggests trust in AI follows trust in the institution, so hospitals need to earn both together.
Patients are not asking whether the algorithm is trustworthy. They are asking whether you are. Every AI tool a clinic introduces is a withdrawal from that account, and the balance was never infinite.
As a patient or clinician, what would make you trust an AI tool used in your care?
